Hardened D&D spell extraction prompts

docs/integrations/dnd-spell-artifacts.md

@@ -50,11 +50,21 @@ Each spell cast contains exactly these required fields:
 - `source_refs`: transcript source references with extractor-assigned source
   IDs and evidence unit ranges. It must contain at least one entry.
 
-All four string fields must be non-empty. `caster` is the in-world caster, not
-the transcript speaker. The `spell` value must resolve through the effective
-SRD-plus-overlay catalog as either a canonical name or alias. Catalog
-validation accepts aliases but does not rewrite them; unknown fields are
-rejected.
+All four string fields must be non-empty. `caster` is the canonical in-world
+caster, not the human player, transcript speaker, or GM when the associated
+character or creature can be identified. Player and party references may
+disambiguate that identity, but do not independently establish that a cast
+occurred. The `spell` value must resolve through the effective SRD-plus-overlay
+catalog as either a canonical name or alias. Catalog validation accepts aliases
+but does not rewrite them; unknown fields are rejected.
+
+`effect` and `narrative_description` record what the transcript establishes
+about the session as played. They do not correct the table from published D&D
+rules or supplement the transcript with model knowledge. When the transcript
+contains a nonstandard or disputed ruling, the artifact may preserve the
+observed resolution and attribute relevant reasoning to the GM or table; it
+must not present that reasoning as a universal game rule. The spell catalog is
+name-recognition policy, not evidence for spell mechanics or outcomes.
 
 ## Source References
 
@@ -63,6 +73,17 @@ Each source reference contains exactly three required fields: `source_id`,
 The unit IDs must be positive integers present in the input, and the start unit
 must not appear after the end unit. Unknown fields are rejected.
 
+For each cast, the complete `source_refs` collection identifies the transcript
+evidence for every factual claim in `caster`, `spell`, `effect`, and
+`narrative_description`. A cast declaration and a later resolution may be
+cited with separate narrow ranges when intervening units are unrelated. A
+reported target, roll, amount, condition, interruption, or outcome must be
+supported by the cited units; otherwise the artifact describes only the
+supported attempt or declaration. The deterministic validators establish that
+ranges are structurally valid and that the spell name is related to cited
+text. Semantic claim completeness is an extraction policy and remains subject
+to evaluation rather than deterministic proof.
+
 Reference slot keys and accepted file types are defined in
 [Configuration](../config.md#implemented-production-modules). References are
 supporting disambiguation material, not source evidence, and are not
@@ -145,3 +166,9 @@ included in manifest metadata. The `normalizer` metadata uses the same catalog
 identity fields when that module is selected. Overlay origin, media type, byte
 size, and raw digest are recorded separately in the manifest's reference
 provenance; see the [JSON output contract](json-output.md#manifestjson).
+
+The extractor's prompt hash, private response-schema hash, and effective catalog
+digest also contribute independently scoped semantic checkpoint fingerprints.
+Changing any of those prepared contracts intentionally produces a cold
+checkpoint miss. Fingerprints contain only digests, never prompt, schema,
+catalog, or reference content.

docs/internal/modules.md

@@ -158,11 +158,25 @@ reference into an immutable effective catalog and adds a generated
 canonical-name-only JSON input to every structured completion request. Overlay
 failures therefore stop construction before source parsing or an LLM call;
 campaign references remain separate disambiguation inputs and never become
-source evidence. Both the extractor and deterministic catalog validator expose
+source evidence.
+
+The prompt requires each cast's source references collectively to support all
+reported facts, using separate narrow ranges when a declaration and resolution
+are non-contiguous. It grounds caster names in in-world identities, using the
+player and party references to disambiguate transcript speakers without
+treating those references as event evidence. Effects describe the session as
+played: model rules knowledge cannot supplement or correct the transcript, and
+nonstandard adjudication is attributed to the GM or table rather than stated as
+a universal rule. Structural source validation remains deterministic; semantic
+claim completeness is enforced through extraction policy and evaluation.
+
+Both the extractor and deterministic catalog validator expose
 the effective base-plus-overlay semantic digest as scoped prepared-component
 checkpoint identity. Raw overlay provenance independently covers file-byte
 changes, while the semantic digest also invalidates reuse when the embedded
-catalog or catalog composition changes. The separate
+catalog or catalog composition changes. The extractor additionally fingerprints
+its complete prompt assets and private response schema, so either semantic
+contract changing invalidates previously recorded extraction checkpoints. The separate
 `internal/modules/dnd/codec/spells` package
 owns the durable schema and stable JSON representation for artifact kind
 `dnd/spell-list`. The runner keeps the result typed through validators and later

internal/modules/dnd/extract/spells/assets/prompts/instructions.md

@@ -2,9 +2,27 @@ Source references must use integer source-unit IDs from the transcript. Provide
 start_unit_id and end_unit_id for each source reference; the extractor assigns
 source_id automatically.
 
+For each artifact, its source references must collectively support every
+factual claim in caster, spell, effect, and narrative_description. Include the
+casting declaration or action and every later transcript unit needed to support
+reported targets, saving throws, damage, healing, conditions, interruptions,
+and outcomes. When supporting evidence is non-contiguous, provide multiple
+narrow source references. Do not use one broad range merely to bridge unrelated
+table conversation. If a detail is not supported by the cited transcript
+units, omit that detail or describe only the supported attempt or declaration.
+
 Return only D&D spell-cast artifacts. For each spell cast, identify the in-world
 caster, spell name, effect, narrative description, and source references.
 
+The caster must be the canonical in-world character or creature, not the human
+player, transcript speaker, or GM. Use the player and party references together
+with transcript context to map first-person player speech to the associated
+player character and use the canonical character name from the references.
+Likewise, attribute a spell narrated by the GM to the in-world creature that
+casts it. Never return a player or GM name when the associated in-world caster
+can be identified. If the caster cannot be resolved, use only the most specific
+in-world identity supported by the transcript; do not invent a name.
+
 Use the canonical spell-name catalog to select spell names. Do not return a
 spell name absent from that catalog, even when it is suggested by general D&D
 knowledge or reference material.
@@ -13,4 +31,11 @@ Use player, party, and glossary reference material only to clarify source text;
 references are not source evidence for a spell cast. Do not return spells,
 casters, or effects that are mentioned only in reference material.
 
+Effects and narrative descriptions are session records, not rules summaries.
+Report only mechanics, explanations, and outcomes established by the cited
+transcript units. Preserve the table's observed resolution without silently
+correcting it from general D&D knowledge. If the transcript gives a possibly
+nonstandard rationale, use wording such as "the GM rules" or "the table
+resolves" rather than asserting that rationale as a universal rule.
+
 Return exactly one JSON object and no explanatory text.

internal/modules/dnd/extract/spells/assets/prompts/task.md

@@ -4,6 +4,13 @@ Extract only spell casts that are supported by the transcript. Do not infer
 spells from general D&D knowledge or from table chatter that does not identify a
 spell being cast.
 
+Describe the session as it was played and adjudicated. The transcript is
+authoritative for what happened in this session, even when a table ruling may
+differ from published D&D rules. Do not correct the transcript or fill in
+unstated mechanics from general D&D knowledge. When a ruling or mechanical
+explanation matters, attribute it to the GM or table instead of presenting it
+as a universal game rule.
+
 Use the provided canonical spell-name catalog when naming each extracted spell.
 Return the canonical catalog spelling exactly. The catalog is a recognition
-aid; it does not establish that a spell was cast.
+aid; it does not establish that a spell was cast or how the spell works.

internal/modules/dnd/extract/spells/assets/schemas/dnd_spells_llm.v1.json

@@ -20,23 +20,28 @@
         "properties": {
           "caster": {
             "type": "string",
-            "minLength": 1
+            "minLength": 1,
+            "description": "Canonical in-world character or creature that casts the spell, never the human player, transcript speaker, or GM when the in-world caster can be identified."
           },
           "spell": {
             "type": "string",
-            "minLength": 1
+            "minLength": 1,
+            "description": "Canonical spell name from the provided spell-name catalog."
           },
           "effect": {
             "type": "string",
-            "minLength": 1
+            "minLength": 1,
+            "description": "Concise effect or resolution established by the cited transcript units; do not infer mechanics from general D&D rules knowledge."
           },
           "narrative_description": {
             "type": "string",
-            "minLength": 1
+            "minLength": 1,
+            "description": "Short session-grounded description containing only details established by the cited transcript units."
           },
           "source_refs": {
             "type": "array",
             "minItems": 1,
+            "description": "One or more narrow transcript ranges that collectively support every factual claim in this spell-cast object, including any reported resolution or outcome.",
             "items": {
               "type": "object",
               "additionalProperties": false,

internal/modules/dnd/extract/spells/extractor.go

@@ -51,6 +51,8 @@ type Extractor struct {
 	llm                contracts.StructuredLLMClient
 	effectiveCatalog   spellcatalog.EffectiveCatalog
 	catalogPromptInput contracts.LLMInputMaterial
+	promptSHA          string
+	responseSchemaSHA  string
 }
 
 func New(llmClient contracts.StructuredLLMClient, _ Options, references ...contracts.ReferenceSet) (*Extractor, error) {
@@ -72,10 +74,20 @@ func New(llmClient contracts.StructuredLLMClient, _ Options, references ...contr
 	if err != nil {
 		return nil, extractorErrorf("prepare spell catalog prompt input: %w", err)
 	}
+	promptSHA, err := scriptoriumPromptMetadata()
+	if err != nil {
+		return nil, extractorErrorf("load prompt metadata: %w", err)
+	}
+	responseSchema, err := loadResponseSchema()
+	if err != nil {
+		return nil, extractorErrorf("load response schema: %w", err)
+	}
 	return &Extractor{
 		llm:                llmClient,
 		effectiveCatalog:   effectiveCatalog,
 		catalogPromptInput: catalogPromptInput,
+		promptSHA:          promptSHA,
+		responseSchemaSHA:  responseSchema.SHA256,
 	}, nil
 }
 
@@ -88,24 +100,18 @@ func (e *Extractor) ReferenceSlots() []contracts.ReferenceSlot {
 }
 
 func (e *Extractor) ManifestMetadata() map[string]any {
-	promptSHA, err := scriptoriumPromptMetadata()
-	if err != nil {
-		promptSHA = ""
-	}
 	metadata := map[string]any{
-		"prompt_id":            PromptID,
-		"prompt_version":       SchemaVersion,
-		"prompt_sha256":        promptSHA,
-		"catalog_base_id":      e.effectiveCatalog.BaseID(),
-		"catalog_digest":       e.effectiveCatalog.Digest(),
-		"catalog_overlay_ids":  e.effectiveCatalog.OverlayIDs(),
-		"response_schema_key":  string(ResponseSchemaKey),
-		"response_schema_id":   ResponseSchemaID,
-		"response_schema_name": ResponseSchemaName,
-	}
-	if schema, err := loadResponseSchema(); err == nil {
-		metadata["response_schema_version"] = schema.Version
-		metadata["response_schema_sha256"] = schema.SHA256
+		"prompt_id":               PromptID,
+		"prompt_version":          SchemaVersion,
+		"prompt_sha256":           e.promptSHA,
+		"catalog_base_id":         e.effectiveCatalog.BaseID(),
+		"catalog_digest":          e.effectiveCatalog.Digest(),
+		"catalog_overlay_ids":     e.effectiveCatalog.OverlayIDs(),
+		"response_schema_key":     string(ResponseSchemaKey),
+		"response_schema_id":      ResponseSchemaID,
+		"response_schema_name":    ResponseSchemaName,
+		"response_schema_version": SchemaVersion,
+		"response_schema_sha256":  e.responseSchemaSHA,
 	}
 	return metadata
 }
@@ -114,7 +120,11 @@ func (e *Extractor) CheckpointFingerprints() []pipeline.CheckpointFingerprint {
 	if e == nil {
 		return nil
 	}
-	return []pipeline.CheckpointFingerprint{{Name: "effective_catalog", Value: e.effectiveCatalog.Digest()}}
+	return []pipeline.CheckpointFingerprint{
+		{Name: "effective_catalog", Value: e.effectiveCatalog.Digest()},
+		{Name: "prompt", Value: e.promptSHA},
+		{Name: "response_schema", Value: e.responseSchemaSHA},
+	}
 }
 
 func (e *Extractor) Extract(ctx context.Context, req contracts.TypedExtractionRequest) (contracts.TypedExtractionResult[dnd.SpellList], error) {

internal/modules/dnd/extract/spells/extractor_test.go

@@ -118,8 +118,18 @@ func TestExtractPromptUsesCanonicalOverlayNamesWithoutAliasesOrMetadata(t *testi
 		t.Fatalf("catalog overlay metadata = %#v", metadata["catalog_overlay_ids"])
 	}
 	fingerprints := newExtractor(t, &fakeSpellsLLMClient{}, overlaySpellCatalogReference()).CheckpointFingerprints()
-	if len(fingerprints) != 1 || fingerprints[0].Name != "effective_catalog" || fingerprints[0].Value != metadata["catalog_digest"] {
-		t.Fatalf("checkpoint fingerprints = %#v, want manifest catalog digest %#v", fingerprints, metadata["catalog_digest"])
+	wantFingerprints := map[string]any{
+		"effective_catalog": metadata["catalog_digest"],
+		"prompt":            metadata["prompt_sha256"],
+		"response_schema":   metadata["response_schema_sha256"],
+	}
+	if len(fingerprints) != len(wantFingerprints) {
+		t.Fatalf("checkpoint fingerprints = %#v, want prompt, response schema, and catalog identities", fingerprints)
+	}
+	for _, fingerprint := range fingerprints {
+		if want, ok := wantFingerprints[fingerprint.Name]; !ok || fingerprint.Value != want {
+			t.Fatalf("checkpoint fingerprint %q = %q, want %#v", fingerprint.Name, fingerprint.Value, want)
+		}
 	}
 	encoded, err := json.Marshal(metadata)
 	if err != nil {